At the end of October Ollama learned to talk: version 0.40.11 added Phonon, a family of small text-to-speech models from Gradium. There is a catch, though: for now speech in Ollama works only on Apple Silicon Macs, with Windows and Linux support announced but not dated. If you are on a PC, or just want to hear it before installing anything, you can try Phonon 31M in your browser.
Phonon 31MGradium · Apache-2.0 (31M English model and voices)A 31M-parameter English text-to-speech model with ten voices, from narrators to a pirate and a dragon: runs on a plain CPU.Try it online →What Phonon 31M is
- 31 million parameters, a 64 MB file in 8-bit GGUF: small enough for a laptop, a phone or a browser tab.
- English, 24 kHz audio, with text normalisation built in: numbers, prices, dates, emails and URLs are read out the way a person would say them.
- A catalog of around a hundred voices, from neutral American and British narrators to characters such as an old pirate, a wizard or a dragon. On LiveWeights we host ten of them.
- It runs on Gradium's native runtime (
ptts), written in Rust: no PyTorch and no GPU needed.
Which Phonon models can be used commercially?
This matters, because the family has two licenses:
| Model | Languages | License |
|---|---|---|
| Phonon 31M EN | English | Apache-2.0: commercial use allowed |
| Phonon 40M (per language) | One language each | CC BY-NC 4.0: non-commercial only |
| Phonon 40M multilingual | Several | CC BY-NC 4.0: non-commercial only |
That is why LiveWeights hosts only the 31M English model: its audio can be used in commercial projects.
How to try it online
- Open the Phonon 31M page and sign in.
- Write up to 600 characters of English text and pick a voice.
- Press run: it runs on our server's CPU and answers in a few seconds. The first run is free.
Try a sentence with numbers and dates (for example a price and a delivery time) to hear the normalisation at work, and one of the character voices for a story.
How it compares
LiveWeights also hosts Paradee 1.2, an even smaller 8M-parameter English model with four voices, and larger GPU models with voice cloning and more languages: Ming 3.0 TTS and Audio8 TTS. We compared them in our text-to-speech roundup.
Then run it locally
The weights are on Hugging Face (gradium/phonon-31m-en). On a Mac with a recent Ollama you can pull the model; on any system you can install Gradium's Python package ptts and generate speech from a script or the command line.